Is it Time to Throw the Data out with the Bathwater?

When it comes to delivering automatic speech recognition (ASR) services, data is a vital part of the food and fuel needed to create accurate systems. But volume isn’t everything – and large volumes of data are certainly not a silver bullet when it comes to creating industry-leading solutions. Carefully selected data ensures diversity of use cases, while also featuring representations of the type of media you expect to be transcribing.

See also: Researchers Surpass Speech Recognition Methods With Help From Honey Bees

Accuracy is affected by several factors such as the quality of the audio equipment, for example, the background noise that might be present when it’s being recorded, and the varying accents and dialects of the people who are speaking. However, all these aspects offer an opportunity for machine learning algorithms to understand these voice characteristics better. So, when it comes to transcribing audio, including these elements, the machine learning algorithms are optimized to deliver best-in-class transcription because they will have been anticipated and built into the training model.

Solutions that depend on huge volumes of data will come under increasing pressure as a result of heightened security surrounding content that includes personal information. This means organizations should think whether even the storage of end-user data for training is the right decision ethically.

Recent market trends suggest that voice is poised to dominate as the next “go-to” user interface. With the recent adoption of voice-enabled smart speakers, speech recognition on almost all smartphones, and the ability to even control your home with your voice, speech is now an everyday method of interacting with technology.

The power of these interactions and ASR has never been so important. ASR can transform interactions into a format that can be used by other products within a solution stack. However, unlike traditional transcription – where low word error rate (WER) was one of the key metrics of success – with the deployment of ASR, even if you get a load of words wrong in the request, if the intent is understood, it’s a successful interaction. When specific industries, organizations, and use cases can potentially have their own unique vocabularies, the flexibility of ASR allows users to adapt their own language models to their specific application. Subsequently, the value of using large quantities of generic speech data seems to be inefficient. It is a case of quality over quantity.

So, just how important is data anyway? If data is the food of machine learning, then – just like the human body – if you want to ensure optimum performance and fitness, the goal is about reducing the amount of low-quality ingredients and replacing them with the best available. Just like with food, smaller amounts of high-quality input will deliver better performance.

Having the knowledge, insight, and skill to make intelligent decisions on the kind of data included within ASR models is crucial to their performance and accuracy. Solely adding huge amounts of voice data won’t give the results. Also, the impact of flooding models with new data, unless carefully managed, can bias models – and, while it might increase accuracy for a specific use case, it can negatively affect others in the long run.